Agile InsightsBriefing
A clockwork movement with a wind-up key, an alarm clock with a blank face and no hands, and a steel ruler with calipers beside a studio microphone and headphones on a dark workbench, labelled Start, Stop, Measure and Episode 7.

Episode 7 · Discussion · About 9 minutes

The Tool That Never Says Stop: AI, ADHD and what an organisation owes its staff

AI helps people with ADHD most with starting work and getting back into it. The same design takes away every cue to stop, and the evidence is thin on both counts.

#people

Download the MP3 · Linda and Leonard, AI voices, read a script written from the sources below.

This episode is the audio version of an essay: The Tool That Never Says Stop, by Gethryn Ghavalas. Read the essay.

In this episode

01How thin the evidence isThe essay's author found no randomised trial of generative AI for adults with ADHD at work, so governance decisions rest on inference. Read with caution, people with ADHD describe AI as helping most with starting and getting back into tasks.
02A chat window has no stopping pointA chat window answers in seconds and never tells you to stop. In one 2025 trial of early tools, experienced developers were nineteen per cent slower with AI yet believed they were faster, so measure session length.
03Running many agents splits attentionA survey, as the essay reports it, found the strain sits in supervising AI, with self reported productivity peaking at about three tools at once. The same tools also help people get back into interrupted work, so setup matters.
04Where the strongest evidence of harm sitsThe essay says the clearest harm comes from AI used on neurodivergent people, not by them. Seven AI writing detectors flagged, on average, 61 per cent of essays by non native English writers as machine written.
05How to configure, protect and measureGive everyone a sanctioned tool without requiring a diagnosis, and never discipline on detector output alone. Configure for stopping, with visible session length and a cap on parallel agents, then measure session length, after hours use and errors alongside satisfaction.
Sources and links (8)
Measuring the impact of early-2025 AI on experienced open-source developers · METR
The randomised trial in which developers were slower with AI but believed they were faster. METR's 2026 update says it no longer reflects current tools.
M365 Copilot evaluation report · UK Department for Business and Trade
The UK pilot that found neurodivergent staff more satisfied with Copilot, and no evidence of higher productivity.
Neurodiversity at work survey · Understood.org / Harris Poll
The 2026 US survey in which 78 per cent of neurodivergent employees said they use AI at work. Press release only.
Hyperfocus in adults with and without ADHD · PubMed
Groen and colleagues (2020) on how common self-reported hyperfocus is among adults in general.
When using AI leads to brain fry · BCG
BCG's 2026 survey on the strain of supervising AI. The detail in the episode rests on secondary coverage.
No task left behind? Examining the nature of fragmented work · University of California, Irvine
Mark, González and Harris (CHI 2005), the observational study behind the 25-minute figure for returning to an interrupted task.
GPT detectors are biased against non-native English writers · Patterns (Cell Press)
Liang and colleagues (2023): seven detectors flagged most of a set of essays by non-native writers as machine-written.
Disability bias in GPT-4 CV screening · ACM Digital Library
Glazko and colleagues (FAccT 2024): a small audit of how GPT-4 ranked CVs that mentioned disability.

Written from public sources and checked on 11 October 2026. This episode is the audio version of a published essay, and every claim comes from that essay as corrected. Some sources could not be opened in full: both Harvard Business Review articles were read from the lead paragraph only, and the ACM pages were checked through arXiv, Crossref and a university repository instead. The BCG detail rests on secondary coverage. Please check the original before quoting. This is general information, not legal or medical advice.

Want to know more?

Sign up for our podcast at agile-insights.com.au.

Read the transcript

Linda AI helps people with A D H D most with starting work and getting back into it, and the same design that helps also takes away every cue to stop. I'm Linda, and this is the Agile Insights discussion series, with the audio version of a published essay. By the end you will know what the evidence says, where it runs out, and what to set up now.

Leonard And I'm Leonard. The question leaders ask is this. We are rolling out AI, so what do we owe our neurodivergent staff? First... the evidence. The author of the essay found no randomised trial of generative AI for adults with A D H D at work.

Linda Not one?

Leonard What exists is thinner: government Copilot trials with volunteers and self reported results, surveys that often have a vendor somewhere in the funding, and a handful of qualitative studies of a dozen or so people. So any governance decision today rests on inference, and a board should be told that.

Linda Read with that caution, the picture is consistent. People with A D H D describe AI as scaffolding for the hardest parts of work: starting, sequencing, and getting back into a task after an interruption. The U K Department for Business and Trade found neurodivergent staff were more satisfied with Copilot than their colleagues. In a twenty twenty six American survey by Understood dot org and Harris, seventy eight per cent of neurodivergent employees said they use AI at work, against fifty nine per cent of everyone else.

Leonard Here I would push back, on behalf of every finance director listening.

Linda I thought you might.

Leonard That same U K pilot found no evidence that the time saved turned into higher productivity. If I am approving licences, that reads like the tool failing. Satisfaction is pleasant, but a business case needs output, and these were volunteers describing their own experience.

Linda The essay argues the measure is wrong... not the tool. Nobody asks glasses or a screen reader to lift output per hour. They exist so a person can work without paying a hidden tax. Producing the same output with less effort, less anxiety and less exhaustion is a gain. The metric counts the output and not what each hour cost.

Leonard What each hour cost.

Linda But you are right on one thing. Satisfaction is not the same as wellbeing, so measure it alongside exhaustion, errors and whether people stay.

Leonard I can accept that. Measure what the hour cost, as well as what it produced. Now the risk. The usual worry is hyperfocus, and that science is shaky. In one large study, eighty three per cent of ordinary adults said they experience it, and diagnosed patients did not clearly differ from matched controls.

Linda Eighty three per cent.

Leonard So the claim that AI makes people with A D H D hyperfocus for hours is not established. What is established is the design. A chat window answers in seconds, never runs out of follow up questions, and has no natural point where it tells you to stop.

Linda The best evidence on time sense comes from a research group called Meter, spelled M E T R. In its twenty twenty five randomised trial, experienced developers were nineteen per cent slower with AI, and afterwards believed it had made them about twenty per cent faster.

Leonard Slower, but feeling faster.

Linda One caveat. That was with early twenty twenty five tools, and Meter's twenty twenty six follow up says the slowdown no longer reflects current ones. What matters is the gap between felt speed and measured speed. For governance, that is the reason to measure session length, and not only ask people how it went.

Leonard The second risk comes from running agents. Not one long session, but a dozen short ones. One first person account, by an agency founder who has never been formally diagnosed with A D H D, describes running twenty agent terminals at once. An eight month study of one technology company found AI intensified work rather than reducing it. So I would set a hard cap on parallel agents for everyone, and call that the answer...

Linda A cap is sensible, but I disagree that it is the answer, because the evidence is more cautious. A survey by B C G located the strain in supervising AI rather than using it. As the essay reports it, fatigue rose with heavy oversight, and self reported productivity peaked at about three AI tools in use at once. Nobody has yet tested whether that strain is larger for workers with A D H D.

Leonard That is fair.

Linda And there is a counterweight. In one much cited observational study, people took about twenty five minutes on average to get back to an interrupted task. A tool that tells you where you were cuts exactly that cost.

Leonard Then we land in the same place. AI can fragment attention... and it can also be the best aid for getting back into work that anyone has built. Which one you get depends on setup. So the cap stays as one setting among several, and configuration becomes a governance decision.

Linda I can live with that.

Leonard The essay says the strongest evidence of harm sits elsewhere, in AI used on neurodivergent people, not by them.

Linda Three findings carry that. Seven AI writing detectors flagged, on average, sixty one per cent of a set of essays by non native English writers as machine written.

Leonard Sixty one per cent.

Linda In a small audit, G P T four usually ranked a C V lower when it added disability related awards and memberships, most sharply for autism and deafness. And since February twenty twenty five, the European Union has banned AI that infers workers' emotions, except for medical or safety reasons. So people are encouraged to use AI to fit in at work... while AI screening marks them down for not fitting in.

Leonard The essay offers three questions that do not wait for the science. One, is the AI compensating or substituting? Help to start and get back into work, with the person still doing the thinking, has the most support. Two, who controls it? A drafting assistant the worker chooses is an accommodation. An engagement score or interview analyser that the employer applies is something else, whatever the vendor calls it.

Linda Whatever the vendor calls it.

Leonard Three, does the configuration put stopping points back? Good setup adds friction where someone should stop, check or switch, and removes it everywhere else.

Linda Now the practical steps. Give everyone a sanctioned tool. Adult A D H D alone is estimated at two to six per cent of Australian adults. In a twenty twenty four U K survey, three in ten neurodivergent employees had not told their manager or H R, and many more have no diagnosis to tell anyone about. So a policy that offers AI only to people who disclose a diagnosis will miss many of the people it exists for.

Leonard Yes. It would.

Linda The Disability Discrimination Act's definition is also broad enough to cover imputed disability, so a diagnosis is not the legal threshold either. Both legal points are general information, not legal advice.

Leonard Next, do not impose a blanket ban without an alternative. Withdrawing a tool that someone has built their working day around is an adjustment question. Never discipline on detector output alone. Keep attention, engagement and sentiment analytics out of individual performance. And configure for stopping, not just for speed. That means short answers by default, one notification when a long agent task is done, visible session length, and a sensible cap on parallel agents.

Linda Then measure it properly. The author is aware of no Australian evaluation that breaks outcomes down by neurotype. Every staged rollout is a natural experiment, so stagger it deliberately, and survey before and during. Screen for traits rather than ask for diagnoses, and keep it voluntary and confidential. Let a university hold the data under ethics approval, so the employer never sees an individual answer. Measure session length, after hours use and error rates alongside satisfaction.

Leonard Here is the standard to hold your AI programme to. The organisations getting this right will not be the ones that bought the most licences. They will be the ones that can say, with data, who the tool is helping, who it is hooking... and what they changed as a result.

Linda If you'd like to know more, sign up for our podcast at agile-insights.com.au.

Leonard Thanks for listening.

More on #people

Episode 34 · Primer and discussionWhen the test stops telling you anything: hunger beats tenureEpisode 32 · Primer and discussionAgentic video is real: what to try, and what to checkEpisode 31 · Primer and discussionSpeeding up planning decisions with AI: who checks the checker?Making You Think 12 · Primer and discussionThe Game Changer, by Jason FoxEpisode 13 · Primer and discussionFairness under pressure: a fair AI decision has to hold when someone leans on itMaking You Think 9 · Primer and discussionArtificial Intimacy, by Sherry Turkle

Every episode tagged #people

All of the briefings (38) · Follow Briefing in Apple Podcasts or any podcast app

Everything from Agile Insights: briefings, Making You Think and the reference series